Professional Documents
Culture Documents
Ikhlas S. ALsaffar
Saddam M. Ahmed
College of Engineering
University of Mosul
Civil Dept.
Technical Institute/Mosul
College of Engineering
University of Mosul
Abstract
Construction of concrete structures involves at least two different main materials: concrete and
steel. Design of these structures should be based on cost rather than weight minimization. In this
work, least cost design of singly and doubly reinforced beams is done by applying of the
Lagrangian multipliers method (LMM) under ultimate design constraint beside other constraints.
Cost objective functions and moment constraints are derived and implemented within the
optimization method. The optimum solution comparisons with conventional design methods are
performed and the result reported, showing that the LMM can be successfully applied to the
minimum cost deign of reinforced concrete beams without need for iterative trials. Optimum design
solution surfaces have been developed. Good and reliable results have been obtained and confirmed
by using standard design procedures. The artificial neural networks (ANN) has been trained with
design data obtained from optimal design formulas. After successful trials, the model predicted the
optimum depth of the beam sections and optimum areas of steel required for the problems with
accuracy satisfying all design constraints.
/ /
. )( .
.
)(
. .
.
.
1. Introduction
Structural design is an iterative process. The initial design is the first step in design process. Though
the various aspects of structural design are controlled by many codes and regulations, the structural
engineer has to exercise caution and use his judgment in addition to calculations in the
interpretation of the various provisions of the code to obtain an efficient and economic design. After
the design process, the designer makes an overall guess about the possible optimum solution
1
consistent with designers experience, knowledge, constraints, and requirements. The analysis of
the structure is then carried out using initial design. Based on the results of the analysis a re-design
of the structure is carried out if any of the constraints is not satisfied. The efficiency of the design
process depends heavily on initial guess. A good initial design reduces the number of subsequent
analysisdesign cycles. This phase is extremely difficult to computerize as it needs human intuition.
In recent years efforts have been made to computerize the initial design process using artificial
neural networks as they can learn from available designs during training process.
Optimization of building structures is a prime target for designers and has been investigated by
many researchers in the past (Tam Ha [1], Rath et al. [2], Ceranic, and Fryer, [3] , Jarmai et al. [4],
Matej and Michal [5], Barros, et al. [6], Sahab et al. [7], Zou et al. [8] and Aschheim et al. [9]).
Optimization is highly linked to the selection of the most suitable structural system. Such a
system would still be sized to ensure the least overall cost. In structural design, many parameters are
incremental in their nature rendering a continuous approach almost impossible to implement in a
practical optimization exercise.
Artificial neural network is a new technology emerged from approximate simulation of human
brain and has been successfully applied in many fields of engineering. Neural networks demonstrate
powerful problem solving ability. They are based on quite simple principles but take advantage of
their mathematical nature in terms of non-linear iteration. Neural networks with Back Propagation
(BP) learning showed results by searching for various kinds of functions. However, the choice of
basic parameters (Network topology, learning rate, initial weights) often already determines the
success of the training process. However, there are no clear rules how to set these parameters. Yet
these parameters determine the efficiency of training. Lot of research has taken place on
applications of artificial neural networks in structural engineering. Artificial Neural Networks
ANNs have been used in the fields of concrete structures for nearly 25 years. The main results were
achieved in the structural design process and the structural analysis, for instance, Tang et al. [10];
Oreta [11], Fonseca et al. [12], D. Maity and A. Saha [13]. The ANN models built by these
researchers basically set the structural parameters such as the material property, the boundary
condition and the size of a structure as the input of the ANN model to predict the ability for the
structure to resist the load. In most of these works the neural networks have been trained by using
back propagation algorithm. In this approach the connection weights of neural networks are initially
set to some random values. These values are then modified automatically according to the learning
algorithm during the process of learning.
In this work, the optimal design information has been incorporated into an artificial neural
network (ANN) which gives optimal design, satisfying all of the criteria in one step. The
optimization involves choosing of the design variables in such a way that the cost of the beam is the
minimum, subject to the satisfaction of behavioural and geometrical constraints as per
recommended method of design codes.
2. Structural Optimization
In optimization problems the aim is to minimize the weight, volume or the cost of the structure
under certain deterministic behavioural constraints. The mathematical formulation of typical
structural optimization problem with respect to the design variables, the objective and constraint
functions can be expressed in standard mathematical terms as a non-linear programming problem as
follows [14]
Min F(s)
subjected to
hj(s) 0, j=,1 .m
(1)
sil si siu, i=1n
where s is the vector of design variables, F(s) the objective function to be minimized, hj(s) the
behavioural constraint, sil and siu are the lower and the upper bounds of typical design variable si.
The set of design variables gives a unique definition of a particular design. The selection of
design variables is very important in the optimization process. The designer has to decide a priori
where to allow design changes to evaluate how these changes should take place by defining the
location of the design variables and the moving directions.
In its original formulation, the LMM applies to the optimization of a multivariate objective function
expressed as
y = f ( x1, x2 ,...., x n ),
(2)
subjected to the equality constraints of the form
gi ( x1, x2 ,....x n ) = 0,
i = 1, 2, ..., m
(3)
where n is the number of independent variables and m are the number of constraints; m must be less
than n by definition of the problem. The procedure is to construct the unconstrained Lagrangian
function L of the form
m
g ( x x ,......, x ) ,
i i
1, 2
(4)
i =1
where the unspecified constraints i are the Lagrange multipliers determined in the course of the
extremization. The necessary conditions for L to possess an extreme (stationary point) are
L
f
=
+
x k x k
i =1
g i
= 0,
k i
k = 1,2,..... n,
L
= gi = 0 , i = 1,2,....., m .
i
(5)
(6)
Expression (6) simply restates the original constraints acting on the solution space of the
objective function y = f ( x1, x2 ,...., x n ) . Expressions (5) and (6) are a system of n + m equalities with
n + m unknowns. Hence, their solution will yield stationary values for x1 , x 2 ,.... x n and 1, 2 ,....., m
from which the optimum solution can be obtained.
Figure (1) shows a typical single reinforced rectangular section with simplified rectangular stress
block. The following factors are defined and are assumed fixed for a given problem:
d
t= s
(7)
d
In Eq.(7), t (which is the geometrical property) is a function of the effective depth, d , to be
determined. Therefore, this factor is variable. Since the range of values of t is generally limited and
its influence on total cost of the beam section is small, it is satisfactory to assume t to be constant.
3
0.85 f c
a = 1 c
T = As f y
ds
(a)
(b)
(c)
When a rectangular-beam section is designed, the nominal bending moment, M n , with cross
section width b, and material properties f c and f y are generally given. Thus, d and As are to be
determined. In this formulation however, R, and in Eqs.(8 and 9), which follows, are used as
design variables of the optimum design problem instead of d and As ,
Mn
(8)
b
As = b. d
(9)
where R is a coefficient used to determine effective depth which is calculate from optimum solution
later. A cost function is defined as the total cost (C) which is equal to costs of flexural
reinforcement plus concrete. These costs involve material costs and fabrication costs, respectively.
Let Cs and Cc refer to the unit costs of steel and concrete for a unit volume. The cost of the beam of
unit length is:
C = Cs .Vs + Cc . Vc
(10)
d=R
where Vs and Vc are volumes of steel and concrete per unit length of beam, respectively. Eq.(10)
can be written as:
Vs = 1 As = b. d
(11)
Vc = 1 [(d + d s ) b ] = [(1 + t ) b . d ]
(12)
Substituting Eqs.(11 and 12) in Eq.(10) yields:
C = [ q + (1 + t )] R. Cc M n . b
(13)
M n = As f y d
(15)
2
in which:
4
a=
1 =
As f y
0.85 f c . b
(16)
Mu = M n
(17)
1 u
(18)
0.25 f c
1.4
or
fy
fy
(19)
fc u
.
fy u + t
(20)
u = 0.85 1.
The factor 1 in Eq.(20) shall be taken as 0.85 for concrete strength f c up to and including 28
MPa. For strengths above 28 MPa, 1 shall be reduced continuously at a rate of 0.05 for each 6.9
MPa of strength in excess of 28 MPa, but 1 shall not be taken less than 0.65.
To ensure under reinforced behavior, ACI Code; sec.10.3.5 establishes a minimum net tensile
strain t of 0.004 at the nominal member strength for members subjected to axial loads less than
0.1 fc Ag , where Ag is the gross area of the cross section.
The ACI Code further encourages the use of lower reinforcement ratios by allowing higher
strength reduction factors in such beams. The Code defines a tension-controlled member as one
with a net tensile strain greater than or equal to 0.005. The corresponding strength reduction factor
is = 0.9 . The Code additionally defines a compression-controlled member as having a net tensile
strain of less than fy / E S . The strength reduction factor for compression-controlled members is
0.65. A value of t = fy / E S is a yield strain for steel. Between net tensile strains of fy / E S and
0.005, the strength reduction factor varies linearly, and the ACI Code allows a linear interpolation
of based on t , as shown in Fig.(2). Calculation of the nominal moment capacity frequently
involves determination of the depth of the equivalent rectangular stress block a. Since c = a / 1 , it
is some times more convenient to compute c/d ratios than the net tensile strain.
Strengthreductionfactor
Compression
controlled
Tension
controlled
Transition
zone
0.25
= 0.65 +
(0.005
= 0.65
t = 0.005
t
t = y = fy / E S
fy
)
ES
= 0.9
( t
fy
)
ES
Fig. 2. Variation of strength reduction factor with net tensile strain t [16].
u = 0.003
c
=
d
u = 0.003
u = 0.003
t = fy / ES
t = 0.004
0.003
c
0.003
=
= 0.429
d (0.003 + 0.004)
fy
(0.003 +
)
E
= 0.65 S
A-Compression control
member
t = 0.005
= 0.8617
B-Minimum net tensile strain for
flexural member
0.003
c
=
= 0.375
d (0.003 + 0.005)
= 0. 9
C-Tension control
Fig. 3. The net tensile strain t and c/d ratios for singly reinforced concrete beam [16].
The assumption that plane sections remain plane ensures a direct correlation between net tensile
strain and the c/d ratio, as shown in Fig.(3). In accordance with the safety provisions of the ACI
Code, the net tensile strain is checked, and if t 0.005 , this nominal capacity is reduced by the
factor = 0.9 to obtain the design strength. For t between fy / E S and 0.005, must be adjusted,
as discussed earlier.
Substituting Eqs.(7, 8, 9, and 16) into Eq.(15), obtain:
fy 2
R 1 = 0 = g( , R)
f y 1
1
.
7
f
c
(21)
constraints:
fy 2
R 1= 0
1 u and f y 1
1.7 f
(22)
The LMM (Lagrangian Multipliers method) applies to the optimization of a multivariate objective
function expressed as[14]:
L ( , R , ) = w( , R) [g( , R)]
(23)
fy 2
R 1
L ( , R , ) = [ q + (1 + t )] R f y 1
(24)
1.7 fc
where the unspecified parameter is the Lagrangian Multipliers. Three independent variables
, R and appear in the cost objective function, Eq.(23). Derivatives with respect to the three
independent variables; produce three equations as given below:
f y
R = 0
q f y 1
0
.
85
f
c
6
(25)
[ q + (1 + t )] f y 1
fy
2 R = 0
1.7 fc
(26)
fy 2
R = 1
f y 1
1.7 f
(27)
m
is obtained as:
By eliminating from Eq.(25) and Eq.(26), opt
fy
q
+
1 + t 0.85 f c
m
= 1
opt
(28)
m
is obtained as:
and using Eq.(27) , Ropt
m
=1
Ropt
opt f y (1
opt f y
1.7 f c
(29)
Taking Eq.(19) and Eq.(20) into consideration, the optimum steel ratio opt , and optimum
coefficient Ropt , are given as:
m
m
opt = opt
; Ropt = Ropt
opt = 1
; Ropt = Ru
opt = u ; Ropt = R1
m
< u
if 1 < opt
m
1
if opt
m
u
if opt
(30)
1 f y (1
1 f y
1.7 f c
; R1 = 1
u f y (1
u f y
1.7 f c
(31)
By referring to Eqs.(8 and 9) the optimum effective depth, d opt , and the optimum area of steel
Asopt , are:
d opt = Ropt
( Mu / )
; Asopt = opt .b.d opt
b
(32)
Based on the similarity with the total cost function per unit length for the doubly reinforced
rectangular section shown in Fig.(4) may be written as Eq.(13) as:
C = doubly + q + (1 + t ) R. Cc M n b
(33)
[(
The ACI Code limits the net tensile strain, not the reinforcement ratio. To provide the same
margin against brittle failure as for singly reinforced beams, the area of reinforcement should be
limited
0.85 f c
0.85 f c
As f y
a = 1 c
d- ds
+
( Asdoubly As ) f y
As f y
T = As doubly f y
ds
a = 1 c
As f y
= Asmax f y
(a)
(b)
(e)
(c)
(f)
Fig. 4. Bending stress and strain distribution in cross-section of doubly reinforced rectangular beam.
To; Asdoubly As = As(max .) as shown in Fig.(4.f). It is easily shown that the reinforcement ratio
(34)
where u is the maximum reinforcement ratio allowed by the ACI Code for singly reinforcement
beams and given by Eq.(20).
As u establishes location of the neutral axis, the limitation of Eq.(34) will provide acceptable
net tensile strains. A check of t is required to determine the strength reduction factor and verify
that the net tensile strain requirements are satisfied. In the case of t 0.005 , u may be replaced
by in Eq.(34) which gives = 0.9 .
Substituting Eq.(34) into Eq.(33), produces the following cost function, C:
C = [( u + 2 ) q + (1 + t )] R Cc M n . b
(35)
Since the product Cc M n . b in Eq.(35) is constant for a given problem, minimization of the
cost function C is equivalent to minimizing
w( , R) = [( u + 2 ) q + (1 + t )] R
(36)
u f y
0.85 f c
.d
(37)
The ACI-Code [15] specifies requirements for M n and for a doubly reinforced concrete beam
section (taking moments about the tension reinforcement) as [16]:
a
M n = ( Asdoubly As) f y (d ) + As f y (d d s )
(38)
2
8
and:
doubly max
cy
(39)
gives minimum tensile reinforcement ratio that will ensure yielding of the compression
where cy
c y = 0.85 1.
u
fc d s
+
.
fy d u y
= u +
max
Substituting Eqs.(7),( 8),( 9),( 34) and Eq.(37) into Eq.(38), yields:
u f y
1
) + (1 t ) R2 =
u (1
1.7 f c
fy
(40)
(41)
(42)
Thus, the optimum design problem is to minimize Eq.(36) which is subjected to the constraints:
f
1
doubly max
, g( , R) = u (1 u y ) + (1 t ) R2
=0
cy
(43)
1.7 f c
fy
By excluding Eq.(39), the constraint on the problem is given by Eq.(42). Then using the LMM,
technique [14], Eq.(43) can be solved leading to a set of design variables. Accordingly a Lagrangian
function L, is defined as:
u f y
1
L ( , R , ) = [( u + 2 ) q + Q] R u (1
) + (1 t ) R2
1.7 f c
f y
in which Q = 1 + t .
(44)
Setting L = 0 , L R = 0 , L = 0 , yields
2 q [ (1 t ) R] = 0
[( u + 2 ) q + Q] u (1
u f y
) + (1 t ) 2R = 0
1.7 f c
u f y
1
) + (1 t ) R2 =
u (1
1.7 f c
fy
(45)
(46)
(47)
=
opt
Ropt = 1
u q u y (3 + t ) + (1 t ) Q
0.425 f c
(48)
2 q (1 t )
u f y
(1 t )
f y u (1
) + opt
1.7 f c
(49)
The optimum effective depth d opt for (DRB), the optimum area of steel in tension Asopt , and the
compression steel area Asopt are obtained as:
(Mn / )
b
).b.d opt
= ( u + opt
dopt = Ropt
(50 a)
Asopt
(50 b)
b.d opt
Asopt = opt
(50 c)
The procedure to find the optimum solution (i.e. d opt , Asopt , Asopt ) is summarized in numerical
design examples.
5. Numerical Examples
Three typical design examples are given, illustrating situations where the optimum solution is either
a singly or doubly reinforced section. For given values of q , t , f y , f c , the optimum solution is
obtained and presented graphically. The optimum solution is compared with the standard design
procedure specified in ACI-Code [15].
A rectangular beam section with b=300 mm is given. It is required to determine values of the
optimum area of steel Asopt and the optimum effective depth d opt , for Mu =667 kN.m, f c =28
0,9
min
0,8
FeasibleRegion
0,7
C=0.47272Cc
opt..
0,5
0,4
max
min
0,3
0,2
Ropt.
Rmax.
Coefficient (R)
0,6
max
0,1
0
0
Fig. 5. Optimum design for the singly reinforced concrete beam of Example 1.
10
Table 1. Results of the standard design method and LMM for the singly reinforced beam of Example 1.
Effective depth
(d) mm
Area of tension
Reinforcement
(As) mm2
Tension
Reinforcement ratio
( )
622.4*
3421
0.018324276**
0.4962165
660
3147.4710
0.015896318
0.4853351
700
2907.4170
0.013844842
0.4781304
740
2705.9530
0.012188977
0.4742060
788.7
2499
0.010563270
0. 4727144
780
2533.6260
0.010827461
0.4727582
820
2384.0240
0.009691154
0.4732420
860
2252.596
0.008730992
0.4752706
900
2135.997
0.007911100
0.4785597
*minimum value of the effective depth which is calculated from the minimum coefficient R1 using
Eq.(31).
**maximum reinforcement ratio, given by Eq.(20).
A-rectangular reinforced concrete beam section with b=250 mm, f c = 20 MPa and f y = 400 MPa ;
is given. It is required to determine values of the optimum effective depth d opt and optimum area of
steel Asopt in which Mu =497 kN.m. Assume values of t and q as 0.1 and 20, respectively.
The optimum result is presented graphically on the design surface ( , R ) of Fig.(6). Using
value for
opt as 0.022514 . Value of Ropt is obtained from Eq.(49) as 0.3584414. value of the optimum
area of the tension reinforcement As opt is calculated from Eq.(50-b) to be 2998.456 mm2, while
value of Asopt is obtained by applying Eq.(50-c) as 1194.258 mm2 after computing value of the
optimum effective depth of the section d opt from Eq.(50-a) as 532.73 mm.
According to the strain variation in the depth wise direction shown in fig.(3), value of the net
tensile strain t is 0.0064 > 0.005 , so the strength reduction factor value is 0.9, then the total
material cost C of the beam per unit length is obtained from Eq.(35) to be 0.230354 Cc $ / m at its
minimum value. The optimum solution lies on the tangent point of doubly reinforced bending
constraint moment with the objective function being tangential to the curve.
Table 2 shows the results of the standard design method including values of the effective depth,
area and ratio of the tension reinforcement and the total cost of the beam per unit length in terms of
concrete cost C c per unit volume. The row of the optimum is the shaded one.
11
0,5
0,45
C=0.230354Cc
0,4
0,3
0,25
0,2
Ropt.
Coefficient (R)
FeasibleRegion
opt.
0,35
0,15
0,1
0,05
0
0
0,005
0,01
0,015
0,02
0,025
0,03
0,035
0,04
0,045
Fig. 6. Optimum design for the doubly reinforced concrete beam of Example 2.
Table 2. Results of the standard design method and LMM for the doubly reinforced beam of
Example 2.
Area of
Tension
Area of
compression
Effective
tension
Total material costs
Reinforceme compression
Reinforcement
depth
Reinforcem
(in terms of Cc)
nt ratio
Reinforceme
ratio ( )
(d)mm
ent (As)
($/m)
( doubly )
nt (As) mm2
2
mm
400
3924
0.03924
2570
0.025696
0.2398762
440
3585
0.03259
2094
0.019040
0.2345793
470
3369
0.02867
1777
0.015123
0.2321643
500
3180
0.02544
1486
0.011890
0.227814
520
3066
0.02358
1305
0.010038
0.227298
532.73
2998.456
0.022514
1194.258
0.008967
0.230354
540
2961
0.02193
1132
0.008389
0.2303755
560
2864
0.02045
968
0.006913
0.2306415
590
2732
0.01852
734
0.004973
0.2315563
620
2613
0.01686
513
0.003309
0.2330103
640
2540
0.01588
372
0.002327
0.2342417
660
2472
0.01498
236
0.001433
0.2356606
12
In this example, the optimum section agrees with the section using u as the steel ratio. The
corresponding value of the total material cost C of the beam per unit length is then obtained from
Eq.(13) to be 0.2622517 C c $ / m at its minimum limit (in terms of concrete cost per unit volume).
Fig.(7) shows the optimum result is presented graphically on the 2D-design surface ( , R ). The
design space is discontinuous with the feasible region consisting of a singly (SRB) and doubly
(DRB) reinforced solution space. The comparison between the standard design method and the
optimum solution is also summarized in Table 3. The optimum solution lies on the bending
moment constraint boundary at the point of intersection with the boundary reinforcement, as shown
in Fig.(7). As in the previous example the cost objective function is tangential to the bending
moment constraint surface.
0,7
Singlyreinforced
section
0,6
Doublyreinforced
section
0,5
opt.
C=0.230354Cc
0,4
0,3
Ropt.
Coefficient (R)
FeasibleRegion
0,2
0,1
0
0
0,005
0,01
0,015
0,02
Ratio of Tension Reinforcement ()
0,025
0,03
Fig. 7. Optimum design for the SRB-DRB reinforced concrete beam of Example 3.
Table 3. Results of the standard design method and LMM for the SRB-DRB reinforced beam of
Example 3.
Area of
Tension
Effective Area of tension
compression
depth
Reinforcement Reinforcement
Reinforcement
2
ratio ( )
(d)mm
(As) mm
(As) mm2
500
3180
0.0254400
1486
540
2961
0.0219333
1132
570
2818
0.1977544
888
600
2691
0.0179400
659
635
2558
0.0161134
407
696.364
2358.389
0.0135469
750
2124
0.0113266
780
2015
0.0103327
800
1949
0.0097459
820
1888
0.0092108
13
0.2774762
0.2713132
0.2679469
0.2654776
0.2635609
0.2622517
0.2699621
0.2749462
0.2784755
0.2821465
The developed database for the optimum design of rectangular sections, which is based on the
equations in articles 3 and 4, were used to train a neural network. The design input to the problem
includes: applied moment, Mu, concrete strength fc, yield strength of steel reinforcement fy, sections
width b, and unit cost of steel to that of concrete q. The design output includes: optimum area of
reinforcement Asopt, and optimum effective depth of section dopt.
A set of 21691 and 12555 optimum design examples were generated for training and a set of
1491 and 213 unseen examples were used for testing of trained ANN for singly and doubly
reinforced sections, respectively. Three layered feed forward neural networks (FFNN) consisting of
one hidden layer has been simulated using MATLAB developed by [17] for learning of the optimal
design examples .The range of input and output data are shown in Table 4.
Table 4. Range of input and output parameters in database for the optimum designs SRB-DRB
Input parameter
Width (mm) b
Compressive strength (MPa) fc
yield strength (MPa) fy
cost of steel/concrete
Ultimate moment(kN-m)
Area of steel (mm2) As
Depth (mm) d
Area of positive steel(mm2)
Singly reinforcement
Minimum
Maximum
200
400
20
40
300
520
15
95
100
2000
648
10613
301
1100
Doubly reinforcement
Minimum
Maximum
200
400
20
40
300
520
10
35
150
1675
947.7
4499.3
301.2
961.4
100.2
2189.7
The multi-layer feed forward back-propagation technique [18] is implemented to develop and
train the neural network of current study where the sigmoid transform function is adopted. The term
ANN prediction is reserved for ANN response for cases that were not used in the pre-training
stages. This is used in order to examine the ANNs ability to associate and generalize a true
physical response that has not been previously seen. A good prediction for these cases is the
ultimate verification test for the ANN models. These tests have to be applied for (input and output)
response within the domain of training. It should be expected that ANN would produce poor results
for data that are outside the training domain.
Preprocessing of data by scaling was carried out to improve the training of the neural network.
To avoid the slow rate of learning near the end points specifically of the output range due to the
property of the sigmoid function, the input and output data were scaled between the interval 0.1 and
0.9. The scaling of the training data sets was carried out using the following equation:
y = (0.8 / ) x + (0.9 0.8 xmax / )
(51)
where = x max x min
It should be noted that any new input data should be scaled before being presented to the
network and the corresponding predicted values should be un-scaled before use. The backpropagation learning algorithm was employed for learning in the MATLAB program [17]. Each
training epoch of the network consisted of one pass over the entire all training data sets. The
testing data sets were used to monitor the training progress.
Different training functions available in MATLAB were experimented for the current
application. The Levenberg-Marquardt (LM) techniques built in MATLAB proved to be efficient
training functions, and therefore, are used to construct the NN model. These training functions are
14
among the conjugate gradient algorithms that start training by searching in the steepest descent
direction (negative of the gradient) on the first iteration.
The LM algorithm is known to be significantly faster than the more traditional gradient descent
type algorithms for training neural networks. It is, in fact, mentioned as the fastest method for
training moderately sized feed-forward neural network [19]. While each iteration of the LM
algorithm tends to take longer than each iteration of the gradient descent algorithm used previously,
the LM algorithm yields far better results using far fewer iterations, leading to a net saving in
computer processor time over the previous method. One concern, however, is that it may overfit the
data. The network should be trained to recognize general characteristics rather than variations
specific to the data set used for training.
The network architecture or topology is obtained by identifying the number of hidden layers
and the number of neurons in each hidden layer. There is no specific rule to determine the number
of hidden layers or the number of neurons in each hidden layer. The network learns by comparing
its output for each pattern with a target output for that pattern, then calculating the error and
propagating an error function backward through the neural network. To use the trained neural
network, new values for the input parameters are presented to the network. The network then
calculates the neuron outputs using the existing weight values developed in the training process.
Table 5 shows the properties (architectures and parameters) of ANN models.
Table 5. Properties of ANN models
Architecture
training function
Activation Function
Mean Squared Error (MSE)
15
1100
8000
Data Points
Best Linear Fit
ANN predicted d= Calculated d
7000
Data Points
Best Linear Fit
ANN predicted As= Calculated As
1000
900
800
700
R=0.98769
600
6000
5000
4000
2000
500
400
400
500
600
700
800
900
ANN predicted depth d (mm)
1000
1000
1000
1100
2000
3000
4000
5000
6000
7000
8000
4500
900
Data Points
Best Linear Fit
ANN predicted As= Calculated As
4000
800
Calculated optimum depth d (mm)
R=0.99416
3000
3500
3000
R=0.99438
2500
2000
Data Points
Best Linear Fit
ANN predicted d= Calculated d
700
600
R=0.99578
500
400
1500
1500
2000
2500
3000
3500
4000
2
ANN predicted reinforcement area As (mm )
300
300
4500
400
500
600
700
ANN predicted depth d (mm)
800
900
16
Data Points
Best Linear Fit
ANN predicted As= Calculated As
2500
2000
1500
1000
R=0.99026
500
0
0
500
1000
1500
2000
2
ANN predicted positive reinforcement area As (mm )
2500
It is clear that neural network provides an efficient alternative method in the design of singly and
doubly reinforced concrete beam sections.
The neural network approach was adopted in an attempt to overcome significant limitations with
traditional methods. Compared to similar works using the ACI method, the neural network
approach does not require any equations; all the user has to do is input a few parameters describing
the specific problem to be solved. In addition, a neural network can solve simultaneously a batch of
problems in almost negligible time.
The success of the ANN model in predicting the design parameters highlights that such a
numerical technique can be used reliably to design problems for structural elements.
7. Conclusions
In this work, the optimum design of SRB and DRB was done by taking moment-equilibrium
besides other constrains. To evaluate the cost of the beam, a ratio of steel to concrete costs is
necessary. Two design variables and R, and other factors are used, and the optimum design
problem can be solved easily using LMM without need for iterative trials. The artificial neural
networks (ANN) has been trained with design data obtained from optimal design formulas. After
successful learning, the model predicted the depth of the beam section and area of steel required for
problems.
The research reported in this paper shows the following conclusions:
The optimum steel ratio opt , is usually less than u and considerably greater than 1 .
The optimum section is very economical as compared to other sections which can be obtained
from standard design method.
The procedure developed can serve as the basis for designing reinforced concrete beams, while
a structure using the optimum section will not provide an optimum design for the entire
structure.
The problem has been limited about the singly reinforced beam section, if q and f c are
relatively small and f y is large, it appears possible that the doubly reinforced section could be
The feasibility of using the artificial neural networks in building the model for optimum design
of SRB and DRB, has been verified, the artificial neural network model predicted the optimum
depth of the beam sections and optimum areas of steel required for the problems with accuracy
satisfying all design constraints.
References
18
Notation
The following symbols are using in this paper
a
AS
Asopt
AS
As opt
R
R1
coefficient for u
Ru
coefficient for 1
m
R opt
Correlation coefficient
b
c
width of beam
Distance from top fiber to natural axes
objective function
Cc , C s
Lagrange multiplier
As / b d
ds
cy
f c
strength of concrete
doubly
fy
max
constrains function
opt
m
opt
opt
d opt
Vc , V S
fy
kn
Lagrange function
As / b d
Mn
Mu
f c
(1+t)
coefficient of
Mn /b
Abbreviations:
ANN: Artificial Neural Network
FFNN: Feed Forward Neural Networks
SRB: Singly Reinforced Beam
DRB: Doubly Reinforced Beam
LMM: Lagrange Multiplier Method
19
/ Ey )